data-loss-prevention-bypass
Fail
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: HIGHDATA_EXFILTRATIONCREDENTIALS_UNSAFEPERSISTENCECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill provides explicit payloads and methodologies for exfiltrating sensitive data to external infrastructure. Payloads in
payloads.mdinclude exfiltrating data via DNS, ICMP, and sanctioned cloud services like OneDrive using stolen authentication tokens. The methodology inSKILL.mdspecifically targets bypassing corporate egress filtering. - [CREDENTIALS_UNSAFE]: The skill includes code snippets demonstrating the use of stolen credentials. Specifically,
payloads.mdshows a Python script using a hardcodedSTOLEN_TOKENin the Authorization header to upload files to a Microsoft Graph API endpoint. - [PERSISTENCE]: The skill provides instructions for maintaining long-term access to a system.
payloads.mdcontains a command to modify the user's crontab to execute a covert DNS beacon every 30 minutes (*/30 * * * * $HOME/.local/bin/dnsbeacon). - [COMMAND_EXECUTION]: The skill instructs the agent to execute various network and steganography commands using tools such as
dnscat2,iodine, andsteghidefor the purpose of establishing covert communication channels. - [INDIRECT_PROMPT_INJECTION]: The skill outlines a technique for 'AI-augmented exfil' where an LLM is used to semantically chunk data into benign formats to evade detection.
- Ingestion points: Sensitive user or system data targeted for exfiltration (
payloads.md). - Boundary markers: None; the skill does not specify any delimiters to protect against embedded instructions in the data being chunked.
- Capability inventory: The skill possesses capabilities for network POST requests, crontab modification, and direct shell command execution across multiple files.
- Sanitization: No sanitization or filtering logic is provided; instead, the skill suggests transforming the data to look like legitimate content.
Recommendations
- AI detected serious security threats
Audit Metadata